Augur

Augur is an agent-agnostic skill package that operationalizes IRIS (ICLR 2025) as a reusable neuro-symbolic SAST workflow.

The package is not tied to any SDK, CLI wrapper, or agent runtime. Any agent that can:

1. Read files
2. Run shell commands
3. Call an LLM

can execute Augur end to end.

What Augur contains

• WORKFLOW.md: the execution contract (two phases, checkpointed workflow)
• references/: CodeQL and framework guidance loaded on-demand
• assets/: reusable schema/template assets (taint + extraction templates)
• scripts/: deterministic helper scripts
• examples/: worked end-to-end example artifacts
• adapters/claude-code/SKILL.md: optional frontend for Claude Code

Scope

Augur focuses on:

• API candidate extraction with CodeQL
• LLM-based source/sink/sanitizer labeling
• CodeQL library/query generation
• Sequential taint analysis runs per CWE pass
• Post-hoc false-positive triage and markdown reporting

Supported languages and frameworks

Languages:

• Python
• Go
• Java
• JavaScript/TypeScript
• C#

Framework references included:

• FastAPI, Django, Flask
• Gin, Echo
• Spring Boot
• Express
• ASP.NET Core

Quick start (any agent)

1. Read WORKFLOW.md.
2. Run Step 0 (pre-flight) before writing or executing queries.
3. Load references/<language>.md and references/<framework>.md based on detection.
4. Execute extraction and taint queries sequentially.
5. Generate labels/labels.json using assets/labels_schema.json.
6. Present checkpoint summary (pause for review in interactive mode).
7. Resume Phase 2, run analysis, then call scripts/filter_and_report.py.

Repository layout

augur/
├── README.md
├── WORKFLOW.md
├── references/
├── assets/
├── scripts/
├── examples/
└── adapters/

Design principles

• Agent-agnostic: no runtime lock-in
• Deterministic shell workflow around probabilistic labeling
• Advisory human checkpoint between inference and exploitation
• Sequential execution for CodeQL DB safety
• Explicit artifact contracts between steps